基于深度学习的篇章级事件抽取研究综述  

Survey on Document-Level Event Extraction Based on Deep Learning

作　　者：胡瑞娟 周会娟[1] 刘海砚[1] 李健[1] HU Ruijuan;ZHOU Huijuan;LIU Haiyan;LI Jian(PLA Strategic Support Force Information Engineering University,Zhengzhou 450001,China)

机构地区：[1]战略支援部队信息工程大学,郑州450001

出　　处：《计算机工程与应用》2022年第24期47-60,共14页Computer Engineering and Applications

摘　　要：事件抽取是信息抽取领域中一项十分重要且具有挑战性的任务,在事理图谱构建、舆情监控、态势感知等方面起着举足轻重的作用。目前研究较多的是句子级事件抽取,而面对“论元分散”和“多事件”的挑战,基于深度学习的篇章级事件抽取陆续展开。总结了篇章级事件抽取的定义、主要任务和面临的挑战,分别从获取词语、句子和文档三种不同粒度的语义信息,捕获文档结构特征建模不同的图结构,融合语义信息和结构特征,以及将事件抽取转化为阅读理解、智能问答等其他任务解决方案等四个不同的维度,详细讨论了近年来篇章级事件抽取相关的模型和方法,在此基础上归纳了常用数据集,并对典型方法的抽取效果进行了评估和对比。展望了篇章级事件抽取的研究趋势。Event extraction is a very important and challenging task in the field of information extraction,which plays a pivotal role in the construction of event evolutionary graph,network monitoring system,and situation awareness.At present,sentence-level event extraction is the most researched task,while document-level event extraction based on deep learning has been created in response to the difficulties of argument-scattering and multi-events,sentence-level.Following a summary of the definition,primary tasksand challenges of document-level event extraction,four distinct dimensions are covered in more detail,including gathering semantic information at various word,sentence,and document granularities,capturing document structural features to model various graph structures,fusing semantic information with structural features,and incorporating event extraction into other task solutions like reading comprehension and intelligent question answer.On the basis of a detailed discussion of the models and techniques connected to domain document-level event extraction,the datasets that are used most frequently are compiled,and the extraction results of common techniques are assessed and contrasted.Finally,document-level event extraction research trends are discussed.

关 键 词：篇章级事件抽取 论元分散 多事件 深度学习 评价指标 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]